In [3]:
%matplotlib inline
import matplotlib.pylab as plt
import numpy as np
import sys
import random
import time
import scipy
import pickle
from keras.optimizers import Adam
import end_to_end_network as etem
from skimage.measure import compare_ssim, compare_psnr
from tqdm import tqdm
Testing model on FastMRI data¶
In [4]:
import cv2 as cv
from scipy import ndimage
import copy
def mse(gt, pred):
""" Compute Mean Squared Error (MSE) """
return np.mean((gt - pred) ** 2)
def nmse(gt, pred):
""" Compute Normalized Mean Squared Error (NMSE) """
return np.linalg.norm(gt - pred) ** 2 / np.linalg.norm(gt) ** 2
def psnr(gt, pred):
""" Compute Peak Signal to Noise Ratio metric (PSNR) """
return compare_psnr(gt, pred, data_range=gt.max())
def ssim(gt, pred):
""" Compute Structural Similarity Index Metric (SSIM). """
return compare_ssim(
gt, pred, multichannel=False, data_range=gt.max()
)
def calculatemetric(imageA, imageB):
return np.array([nmse(imageA, imageB),ssim(imageA, imageB),psnr(imageA, imageB)])
def compare_images(imageA, imageB, imageC, normalize=False):
nimageB = copy.deepcopy(imageB)
if normalize:
cv.normalize(imageB, nimageB, 0, 255, cv.NORM_MINMAX)
mb = nmse(imageA, nimageB)
sb = ssim(imageA, nimageB)
pb = psnr(imageA, nimageB)
mc = nmse(imageA, imageC)
sc = ssim(imageA, imageC)
pc = psnr(imageA, imageC)
imageA = ndimage.rotate(imageA, -90)
imageB = ndimage.rotate(imageB, -90)
imageC = ndimage.rotate(imageC, -90)
fig = plt.figure(figsize=(10,10))
plt.suptitle("Target vs network MSE: %.5f, SSIM: %.2f, PSNR: %.2f" % (mb, sb, pb)+" Target vs Zeroimputed MSE: %.5f, SSIM: %.2f, PSNR: %.2f" % (mc, sc, pc))
ax = fig.add_subplot(2, 2, 1)
plt.imshow(imageA, cmap="gray")
plt.axis("off")
ax = fig.add_subplot(2, 2, 2)
plt.imshow(imageB, cmap="gray")
plt.axis("off")
ax = fig.add_subplot(2, 2, 3)
plt.imshow(imageA, cmap="gray")
plt.axis("off")
ax = fig.add_subplot(2, 2, 4)
plt.imshow(imageC, cmap="gray")
plt.axis("off")
plt.show()
def standardize(maskedkspace):
output = copy.deepcopy(maskedkspace)
temp = np.array(maskedkspace[:,:,:,0]) +1j*np.array(maskedkspace[:,:,:,1])
mean = temp.mean()
std = temp.std()
temp = (temp-mean)/std
output[:,:,:,0] = temp.real
output[:,:,:,1] = temp.imag
return output
def imagetokspace(image):
s = image.shape
kspace = np.zeros((s[0],s[1],2))
temp = np.fft.fft2(image[:,:,0])
kspace[:,:,0] = temp.real
kspace[:,:,1] = temp.imag
return kspace
Loading Testing Data¶
In [5]:
#Testing data
with (open('Xtest.pickle', "rb")) as openfile:
mi_test = np.abs(pickle.load(openfile))
In [6]:
#Testing data
with (open('Xtest.pickle', "rb")) as openfile:
mi_test = np.abs(pickle.load(openfile))
mk_test = np.zeros((mi_test.shape[0],320,320,2))
for i in range(mi_test.shape[0]):
mk_test[i,:,:,:] = imagetokspace(mi_test[i])
In [7]:
import pickle
import numpy as np
with (open('Ytest.pickle', "rb")) as openfile:
oi_test = np.abs(pickle.load(openfile))
In [8]:
print("Total number of data points in test dataset:",mk_test.shape[0])
Results for image constructed from zero imputed/undersampled k-space¶
In [25]:
#Average NMSE, SSIM and PSNR across testing data
results = np.array([0.0,0.0,0.0])
for i in tqdm(range(oi_test.shape[0])):
results += calculatemetric(oi_test[i,:,:,0],mi_test[i,:,:,0])
results/oi_test.shape[0]
Out[25]:
Testing end to end model having k-space MSE and image MSE¶
In [31]:
import os
stats = np.load("training_data_stats20.npy")
model = etem.etem(stats[0],stats[1],stats[2],stats[3],kshape = (5,5),kshape2=(3,3))
opt = Adam()
model.compile(loss = [etem.nrmse,etem.nrmse],optimizer=opt, loss_weights=[0.02, 0.98])
model_name = "end_to_end_MSE.hdf5"
if os.path.isfile(model_name):
print("loading weights")
model.load_weights(model_name)
In [32]:
pred = np.zeros((mk_test.shape[0],320,320,1))
for i in tqdm(range(789)):
index = i*9
pred[index:index+9,:,:,:] = model.predict(mk_test[index:index+9,:,:,:])[1].astype(np.float64)
In [33]:
#Average NMSE, SSIM and PSNR across testing data
results = np.array([0.0,0.0,0.0])
for i in tqdm(range(oi_test.shape[0])):
results += calculatemetric(oi_test[i,:,:,0],pred[i,:,:,0])
results/oi_test.shape[0]
Out[33]:
Looking at some of the results¶
In [35]:
for i in range(10):
compare_images(oi_test[i,:,:,0],pred[i,:,:,0],mi_test[i,:,:,0])
Testing end to end model having k-space MSE, image MSE and image SSIM loss¶
In [21]:
import os
stats = np.load("training_data_stats20.npy")
model = etem.ete(stats[0],stats[1],stats[2],stats[3],kshape = (5,5),kshape2=(3,3))
opt = Adam()
model.compile(loss = [etem.nrmse,etem.nrmse,etem.ssimloss],optimizer=opt, loss_weights=[0.02, 0.80, 0.18])
model_name = "end_to_end_MSE_SSIM.hdf5"
if os.path.isfile(model_name):
print("loading weights")
model.load_weights(model_name)
In [22]:
pred = np.zeros((mk_test.shape[0],320,320,1))
for i in tqdm(range(789)):
index = i*9
pred[index:index+9,:,:,:] = model.predict(mk_test[index:index+9,:,:,:])[1].astype(np.float64)
In [24]:
#Average NMSE, SSIM and PSNR across testing data
results = np.array([0.0,0.0,0.0])
for i in tqdm(range(oi_test.shape[0])):
results += calculatemetric(oi_test[i,:,:,0],pred[i,:,:,0])
results/oi_test.shape[0]
Out[24]:
Looking at some of the results¶
In [23]:
for i in range(10):
compare_images(oi_test[i,:,:,0],pred[i,:,:,0],mi_test[i,:,:,0])
